You are on page 1of 7

1

SOLUTIONS TO CHAPTER 12 MP REVIEW PROBLEMS


1. f''(x) = e-x 0 for all x. Thus f(x) is a convex function. 2. Let x = location of store. We wish to choose x to minimize f(x) = (x - 3)2 + (x - 4)2 + (x - 5)2 + (x - 6)2 + (x - 17)2 f'(x) = 2[ x - 3 + x -4 + x - 5 + x - 6 + x - 17] = 0 for 3 + 4 + 5 + 6 + 17 x = --------------------- = 7. now f''(x) = 10>0 5 so x = 7 is a local minimum. Also f''(x)>0 implies that f(x) is convex so x = 7 does indeed minimize f(x). In general the store should be located at the arithmetic mean of the location of the n customers (i.e. x = (x1 + x2 + ... xn)/n). 3a. Let R = units of raw material purchased. We wish to solve max z = (49 - x1)x1 + (30 - 2x2)x2 - 5R st (1) x1 2R (2) x2 R, All variables 0 Since the objective function is concave and the constraints are linear the K-T conditions will yield an optimal solution. The K- T conditions are (3) 49 - 2x1 - 1 = 0 (4) 30 - 4x2 - 2 = 0 (5) -5 + 21 + 2 = 0 (6) 1(x1 - 2R) = 0 (7) 2(x2 - R) = 0 x1, x2, 1, 2 0 Let's try 1>0 and 2 = 0. Then (6) yields x1 = 2R. From (5) 1 = 2.5. Then (3) yields x1 = 23.25 and (4) yields x2 = 7.5. Since x1 = 2R we find that R = 11.625. All K-T conditions and original constraints are satisfied so we have found an optimal solution. Since the last unit of raw material purchased must have generated $5 in extra revenue, an additional unit of raw material would be worth slightly less than $5. 3b. R = 11.625, x1 = 23.25, x2 = 7.5. 4. For any two points on the curve y = |x| the line joining the two points is never below the curve. Thus |x| is a convex function of x. 5. x1 = 5 - .618(5) = 1.91, x2 = 0 + .618(5) = 3.09 f(x1) = 2.08>f(x2) = -.28 so new interval of uncertainty is [0, 3.09). x3 = 3.09 - .618(3.09) = 1.18, x4 = 1.91 f(x3) = 2.15 >f(x4) = 2.08 so new interval of uncertainty is [0, 1.91). x6 = 1.18, x5 = 1.91 - (.618)(1.91) = .73. f(x6) = 2.15>f(x5) = 1.66 so new interval of uncertainty is (.73, 1.91]. x7 = 1.18 and x8 = .73 + (.618)1.18 = 1.46.

2 f(x8) = 2.25>f(x7) = 2.15 so new interval of uncertainty is (1.18, 1.91]. Now x9 = 1.46 and x10 = 1.18 +.618(.73) = 1.63 and f(x9) = 2.25>f(x10) = 2.23 so new interval of uncertainty is [1.18, 1.63). This interval has width less than .50, so we are finished. (actual maximum occurs for x = 1.5) 6. f(x1, x2) =[exp(-x1-x2)(1 -x1-x2)-1 exp(-x1-x2)(1 -x1-x2)]. (0,1) = [-1 0]. Thus new point is (-t, 1) where t 0 maximizes f(t) = (1 - t)et -1 + t. f'(t) = (1 - t)et-1 - et-1 + 1 = 0 for 1 = te t-1 or t = 1. thus new point is (-1,1) f(-1, 1) = [0 1] so new point is [-1 1+t] and we choose t 0 to maximize h(t) = te-t + 1, h'(t) = -te-t + e-t = 0 for t = 1. Thus new point is (-1, 2). 7. Let xt = number of units produced during month t. Then we wish to solve min x12 + x22 + x32 st x1 + x2 + x3 = 60 L = x12 + x22 + x32 + (60 -x1 - x2 - x3) L/x1 = 2x1 - = 0 , L/x2 = 2x2 - = 0, L/x3 = 2x3 - = 0 L/ = 60 - x1 - x2 - x3 = 0 These equations are satisfied by x1 = x2 = x3 = 20, = 40. Since the objective function is convex and the constraint is linear, this is the optimal solution to the problem. In general, if the cost of producing x units during a month is c(x), we must solve min c(x1) + c(x2) + c(x3) st. x1 + x2 + x3 = 60 L = c(x1) + c(x2) + c(x3) +(60 - x1 - x2 - x3) L/x1 = c'(x1) - = 0, L/x2 = c'(x2) - = 0, L/x3 = c'(x3) - = 0 L/ = 60 - x1 - x2 - x3 = 0 These conditions imply that c'(x1) = c'(x2) = c'(x3) = . If c''(x)>0, then x1 = x2 = x3 must hold. 8. If we choose to maximize ln x + ln y + ln w, then the Lagrangian is L = ln x + ln y + ln w + (36 - 2x - 3y - 4w) L/x = 1/x - 2 = 0, L/y = 1/y - 3 = 0, L/w = 1/w - 4 = 0, (1) L/ = 36 - 2x - 3y - 4w = 0. Thus x = 1/2, y = 1/3, w = 1/4. Then (1) yields 3/ = 36 or = 1/12. Then we obtain x = 6, y = 4, w = 3, optimal objective function value of 72. 9. KT-Conditions yield (1) -50/x2 + y - 1 = 0 (2) -20/y2 + x - 2 = 0 (3) 1(x - 1) = 0 (4) 2(y - 1) = 0, x 1, y 1, 1 0, 2 0. Case I: Trying 1 and 2>0 does not work. Case II: Trying 1>0 and 2 = 0 violates (1). Case III: Trying 1 = 0 and 2>0 violates (2). Case IV: 1 = 2 = 0. Then (1) and (2) yield x2y = 50 and y2x = 20. Solving yields y = 2 and x = 5, which satisfies the KT- conditions and has z = 30. Since this is the only point satisfying KT-conditions, it must be the optimal solution.

3 10. We want to maximize = (10,000 + 5a1/2 - 100p)(p - 10) - a. /p = -100(p - 10) + 10,000 + 5a1/2 - 100p = 0 /a = 5(p - 10)/2a1/2 - 1 = 0 Solving these equations yields p = 58 and a = 14,400. 11. We want to maximize P = 150L1/3M2/3 - 50L - 100M P/L = 50L-2/3M2/3 - 50 = 0. P/M = 100L1/3M-1/3 - 100 = 0. Solving these equations yields L = M. Substituting this in the objective function shows that whenever L = M, P = 0. 12. We want to max H(t) = e-rtF(t). H'(t) = F(t)(-re-rt) + e-rtF'(t) = 0 if r = F'(t)/F(t) Also (1) H'(t)>0 if F'(t)/F(t)>r. Also note that g(t) = F'(t)/F(t) is a decreasing function of t with g(0)>r and g(infinity)<r. This implies that g(t*) = r has a unique solution, and from (1), t* must maximize H(t). 13. We want to show that q maximizes g(p), where g(p) = q[1 (1-p)2] + (1 - q)(1 - p2). Then g'(p) = 2q(1-p) + (1-q)(-2p) = 0 for q = p. Since g''(p) = -2q - 2(1-q)<0, we know that choosing p = q will maximize g(p), as desired. 14. Using the hint we solve max (lnx/x) st -x -a. The K-T conditions yield x(1/x) - ln x -------------- + = 0 , (a-x) = 0, x a, 0. x2 We find the unique solution to the K-T conditions occurs for x = a, and = (ln a - 1)/a2. Thus for b a e we know that (ln b)/b< (ln a)/a or a(ln b)<b(ln a). Taking e to both sides we obtain ba<ab. To solve the problem now let a = e and b = . 15. Let r = radius of desired circle, and (x,y) = coordinates of the desired circle's center. Then we wish to solve min z = r st (0 - x)2 + (0 - y)2 r2 (1 - x)2 + (1 - y)2 r2 (2 - x)2 + (3 - y)2 r2. Using to solve this NLP yields r = 1.8(or (3.25)1/2), x = 1, and y = 1.5. 16. Suppose we produce x units during month 1 and 40 - x units during month 2. We want to show that min x1/2 + (40 - x)1/2 = f(x) 0 x 40 occurs for x = 0 or x = 40. f'(x) = x-1/2/2 - (40 - x)-1/2/2 = 0 for x = 20. However,

4 f''(x) = -x /4 - (40 - x) /4<0, so x = 20 is a local maximum. Since x = 0 and x = 40 are endpoints they may solve the problem. Since for x near 0, f'(x)>0 and for x near 40, f'(x)<0 both x = 0 and x = 40 are local minima. f(0) = f(40) = 40 1/2, so x = 0 and x = 40 are both solutions to the problem. In general, if c(x) = cost of producing x units per month is an increasing concave function we wish to solve (1) min c(x) + c(40 - x) = f(x) 0 x 40 Possible solutions to this problem occur where x = 0, x = 40, and where f'(x) = 0. f'(x) = c'(x) - c'(40 - x) , f''(x) = c''(x) +c''(40 - x)<0. Thus point where f'(x) = 0 cannot be a local min and optimal solution must occur where x = 0 or x = 40. Since f'(0)>0 and f'(40)<0, we know that both x = 0 and x = 40 are local minima. They both have a total cost of c(40), so the optimal solution to (1) is x = 0 or x = 40.
-3/2 -3/2

17a. Let's assume f''(x)>0 for all x. Possible solutions to the NLP occur where x = a, x = b, or at x* satisfying f'(x*) = 0. But f''(x*)>0, so x* cannot be a local max, and optimal solution to NLP must occur for x = a or x = b. A more general argument follows: Suppose f'(a) 0. Then for some x*, f will be non-increasing for x x* and non-decreasing for x>x*. This implies that x = a or x = b must be optimal. Suppose f'(a)>0. Then f(x) is strictly increasing for all x>a, and x = b will be optimal. 17b. Suppose that a<x*<b is optimal for the NLP. Now for some c (0<c<1) we may write x* = ca + (1 - c)b and convexity of f(x) yields f(x*) cf(a) + (1 - c)f(b). This implies that either f(a) f(x*) or f(b) f(x*). This means that if x* solves the NLP, than so does a or b. 18. Let x = location of the store. We wish to choose x to minimize f(x) = |x - 3| + |x - 4| + |x - 5| + |x - 6| + |x - 17| From Problem 4 (and the fact that the sum of convex functions is convex) we know that f(x) is a convex function. Thus any local minimum of f(x) will minimize f(x). We claim that choosing x = 5 (the median of the customer's locations) yields a local minimum. To see this suppose that you move the store location units to the left of 5 ( is small). Then customers 3 and 4 travel less units to store but customers 5, 6, and 17 travel units more to store, so total travel time increases. Suppose you move the store location units to the right of 5. Then customers 3,4 and 5 travel more units to the store while customers 6 and 17 travel less units to the store, and again total travel time increases. Thus x = 5 is a local minimum, and hence a minimum over all possible store locations. In general, if there are n customers, the store should be located at the median value of the customer's locations. If there are an even number (say 2n) of customers, locate the store anywhere between the n'th and n + 1 customer. 19a. K-T conditions are (ignoring non-negativity) original constraints

5 and (1) (2) (3) (4) (5) x1p1'(x1) + p1(x1) - 1 = 0 x2p2'(x2) + p2(x2) - 2 = 0 -c + k11 + k22 = 0 1(x1 - k1z) = 0 2(x2 - k2z) = 0 1 0, 2 0

19b. Now firm wishes to maximize _ f(x1, x2)) = x1p1(x1) + x2p2(x2) - 1x1 - 2x2 _ (1)' f/x1 = x1p1'(x1) + p1(x1) - 1= 0 _ (2)'f/x2 = x2p2'(x2) + p2(x2) - 2 = 0 Clearly any solution to original problem will satisfy (1)' and (2)'. Also the "revenues" in f(x1, x2) are identical to the Revenues in (1)' and (2)'. Can we show that the costs in f(x 1, _ _ _ _ _ x2) (given by 1x1 + 2x2) are identical to the cost cz incurred in the original problem? _ _ _ _ _ _ Case 1: 1>0, 2>0: Then x1 = k1z and x2 = k2z and from (3) total cost is _ _ _ _ _ _ _ _ _ cz = k1z 1 + k2z2 = 1x1 + 2x2, and both problems incur the same production cost. _ _ Case 2: 1 = 2 = 0: From (3) this cannot occur unless c = 0. _ _ _ _ Case 3: 1>0, 2 = 0: Then (3) and (4) yield x1 = k1z and _ k11 = c. Now total cost is _ _ _ _ _ _ _ cz = k11z = x11 + x22 , as desired. _ _ Case 4: 1 = 0, 2>0: The reasoning here is identical to Case 3. In Problems 20 and 21 we use the following idea: suppose we omit several constraints and solve an NLP. If our answer is feasible in the original NLP, then our answer is an optimal solution to the original LP. 20. We temporarily ignore the constraints that the lengths of the three sides must satisfy the triangle inequality. Then the values of a, b, and c may be found by solving max z = s(s - a)(s - b)(s - c) st a + b + c = 60 We choose to max (ln z) = max ln s + ln (s - a) + ln (s - b) + ln(s - c). Using s = (a + b + c)/2 we find our Lagrangian to be _

6 L = ln([(a + b+ c)/2] + ln[(b + c - a)/2] + ln[(a + c - b)/2] + ln[(a + b - c)/2] + (60 - a - b - c). L/a = 1/(a+b+c) - 1/(b+c-a) + 1/(a+c-b) + 1/(a+b-c) - = 0 L/b = 1/(a+b+c) + 1/(b+c-a) - 1/(a+c-b) + 1/(a+b-c) - = 0 L/c = 1/(a+b+c) + 1/(b+c-a) + 1/(a+c-b) - 1/(a+b-c) - = 0 L/ = 60 - a - b - c = 0. The symmetry of the first three equations suggests setting a = b = c = 20. This yields = 1/15 and works. Since our solution satisfies nonnegativity and triangle inequality constraints, it must yield maximum area. Thus maximum area is obtained with (unsurprisingly!) an equilateral triangle. 21. Begin by ignoring the constraints that I p1 p2 F; we'll see that the answer to the unconstrained problem satisfies these constraints, so it must be optimal for the constrained problem. min z = (p1/I)1/2 + (p2/p1)1/2 + (F/p2)1/2 = G(p1, p2). G/p1 = 1/2(p1I)1/2 - (p2)1/2/2(p1)3/2 = 0 if p1 = (Ip2)1/2 (1) G/p2 = 1/2(p1p2)1/2 - (F)1/2/2(p2)3/2 = 0 for p2 = (p1F)1/2 (2). Substituting (2) into (1) yields (3) p1 = I2/3F1/3. Putting (3) into (1) yields (4) p2 = I1/3F2/3. From (3) and (4) we see that I p1 p2 F, so we know our answer solves the problem. 22. If we move away from a point v by the vector d = [d1 d2 ... dn] the (f(v)/x1)d1 + (f(v)/x2)d2 + ... Thus we wish to solve the following i = n i = 1 a distance in a direction specified increase in f is given by + (f(v)/xn)dn). NLP:

max

(f(v)/xi)di

st d12 + d22 + ... dn2 = 1 Forming the Lagrangian we obtain

L =

i = n (f(v)/xi)di + (1 - d12 - d22 - ... dn2) i = 1

(1) L/di = f/xi(v) - 2di = 0 (2) L/ = 1 - d12 - d22 - ... -dn2 = 0 From (1) we obtain (3) di = (f(v)/xi)/2. Substituting (3) into (2) yields ---------------------/ i = n = / (f(v)/xi)2 \/i = 1 ----------------------f(v)/xi

7 2 Now (3) yields di = --------

|| f(v)|| Thus the vector [d1 d2 ... dn] yielding the maximum increase in f when moving away from v has the same direction as the gradient of f at v.

You might also like